PeerStreamer

/*

 * Windows Media Audio Voice decoder.

 * Copyright (c) 2009 Ronald S. Bultje

 *

 * This file is part of FFmpeg.

 *

 * FFmpeg is free software; you can redistribute it and/or

 * modify it under the terms of the GNU Lesser General Public

 * License as published by the Free Software Foundation; either

 * version 2.1 of the License, or (at your option) any later version.

 *

 * FFmpeg is distributed in the hope that it will be useful,

 * but WITHOUT ANY WARRANTY; without even the implied warranty of

 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU

 * Lesser General Public License for more details.

 *

 * You should have received a copy of the GNU Lesser General Public

 * License along with FFmpeg; if not, write to the Free Software

 * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA

 */

/**

 * @file

 * @brief Windows Media Audio Voice compatible decoder

 * @author Ronald S. Bultje <rsbultje@gmail.com>

 */

#include <math.h>

#include "avcodec.h"

#include "get_bits.h"

#include "put_bits.h"

#include "wmavoice_data.h"

#include "celp_math.h"

#include "celp_filters.h"

#include "acelp_vectors.h"

#include "acelp_filters.h"

#include "lsp.h"

#include "libavutil/lzo.h"

#include "avfft.h"

#include "fft.h"

#define MAX_BLOCKS           8   ///< maximum number of blocks per frame

#define MAX_LSPS             16  ///< maximum filter order

#define MAX_LSPS_ALIGN16     16  ///< same as #MAX_LSPS; needs to be multiple

                                 ///< of 16 for ASM input buffer alignment

#define MAX_FRAMES           3   ///< maximum number of frames per superframe

#define MAX_FRAMESIZE        160 ///< maximum number of samples per frame

#define MAX_SIGNAL_HISTORY   416 ///< maximum excitation signal history

#define MAX_SFRAMESIZE       (MAX_FRAMESIZE * MAX_FRAMES)

                                 ///< maximum number of samples per superframe

#define SFRAME_CACHE_MAXSIZE 256 ///< maximum cache size for frame data that

                                 ///< was split over two packets

#define VLC_NBITS            6   ///< number of bits to read per VLC iteration

/**

 * Frame type VLC coding.

 */

static VLC frame_type_vlc;

/**

 * Adaptive codebook types.

 */

enum {

    ACB_TYPE_NONE       = 0, ///< no adaptive codebook (only hardcoded fixed)

    ACB_TYPE_ASYMMETRIC = 1, ///< adaptive codebook with per-frame pitch, which

                             ///< we interpolate to get a per-sample pitch.

                             ///< Signal is generated using an asymmetric sinc

                             ///< window function

                             ///< @note see #wmavoice_ipol1_coeffs

    ACB_TYPE_HAMMING    = 2  ///< Per-block pitch with signal generation using

                             ///< a Hamming sinc window function

                             ///< @note see #wmavoice_ipol2_coeffs

};

/**

 * Fixed codebook types.

 */

enum {

    FCB_TYPE_SILENCE    = 0, ///< comfort noise during silence

                             ///< generated from a hardcoded (fixed) codebook

                             ///< with per-frame (low) gain values

    FCB_TYPE_HARDCODED  = 1, ///< hardcoded (fixed) codebook with per-block

                             ///< gain values

    FCB_TYPE_AW_PULSES  = 2, ///< Pitch-adaptive window (AW) pulse signals,

                             ///< used in particular for low-bitrate streams

    FCB_TYPE_EXC_PULSES = 3, ///< Innovation (fixed) codebook pulse sets in

                             ///< combinations of either single pulses or

                             ///< pulse pairs

};

/**

 * Description of frame types.

 */

static const struct frame_type_desc {

    uint8_t n_blocks;     ///< amount of blocks per frame (each block

                          ///< (contains 160/#n_blocks samples)

    uint8_t log_n_blocks; ///< log2(#n_blocks)

    uint8_t acb_type;     ///< Adaptive codebook type (ACB_TYPE_*)

    uint8_t fcb_type;     ///< Fixed codebook type (FCB_TYPE_*)

    uint8_t dbl_pulses;   ///< how many pulse vectors have pulse pairs

                          ///< (rather than just one single pulse)

                          ///< only if #fcb_type == #FCB_TYPE_EXC_PULSES

    uint16_t frame_size;  ///< the amount of bits that make up the block

                          ///< data (per frame)

} frame_descs[17] = {

    { 1, 0, ACB_TYPE_NONE,       FCB_TYPE_SILENCE,    0,   0 },

    { 2, 1, ACB_TYPE_NONE,       FCB_TYPE_HARDCODED,  0,  28 },

    { 2, 1, ACB_TYPE_ASYMMETRIC, FCB_TYPE_AW_PULSES,  0,  46 },

    { 2, 1, ACB_TYPE_ASYMMETRIC, FCB_TYPE_EXC_PULSES, 2,  80 },

    { 2, 1, ACB_TYPE_ASYMMETRIC, FCB_TYPE_EXC_PULSES, 5, 104 },

    { 4, 2, ACB_TYPE_ASYMMETRIC, FCB_TYPE_EXC_PULSES, 0, 108 },

    { 4, 2, ACB_TYPE_ASYMMETRIC, FCB_TYPE_EXC_PULSES, 2, 132 },

    { 4, 2, ACB_TYPE_ASYMMETRIC, FCB_TYPE_EXC_PULSES, 5, 168 },

    { 2, 1, ACB_TYPE_HAMMING,    FCB_TYPE_EXC_PULSES, 0,  64 },

    { 2, 1, ACB_TYPE_HAMMING,    FCB_TYPE_EXC_PULSES, 2,  80 },

    { 2, 1, ACB_TYPE_HAMMING,    FCB_TYPE_EXC_PULSES, 5, 104 },

    { 4, 2, ACB_TYPE_HAMMING,    FCB_TYPE_EXC_PULSES, 0, 108 },

    { 4, 2, ACB_TYPE_HAMMING,    FCB_TYPE_EXC_PULSES, 2, 132 },

    { 4, 2, ACB_TYPE_HAMMING,    FCB_TYPE_EXC_PULSES, 5, 168 },

    { 8, 3, ACB_TYPE_HAMMING,    FCB_TYPE_EXC_PULSES, 0, 176 },

    { 8, 3, ACB_TYPE_HAMMING,    FCB_TYPE_EXC_PULSES, 2, 208 },

    { 8, 3, ACB_TYPE_HAMMING,    FCB_TYPE_EXC_PULSES, 5, 256 }

};

/**

 * WMA Voice decoding context.

 */

typedef struct {

    /**

     * @defgroup struct_global Global values

     * Global values, specified in the stream header / extradata or used

     * all over.

     * @{

     */

    GetBitContext gb;             ///< packet bitreader. During decoder init,

                                  ///< it contains the extradata from the

                                  ///< demuxer. During decoding, it contains

                                  ///< packet data.

    int8_t vbm_tree[25];          ///< converts VLC codes to frame type

    int spillover_bitsize;        ///< number of bits used to specify

                                  ///< #spillover_nbits in the packet header

                                  ///< = ceil(log2(ctx->block_align << 3))

    int history_nsamples;         ///< number of samples in history for signal

                                  ///< prediction (through ACB)

    /* postfilter specific values */

    int do_apf;                   ///< whether to apply the averaged

                                  ///< projection filter (APF)

    int denoise_strength;         ///< strength of denoising in Wiener filter

                                  ///< [0-11]

    int denoise_tilt_corr;        ///< Whether to apply tilt correction to the

                                  ///< Wiener filter coefficients (postfilter)

    int dc_level;                 ///< Predicted amount of DC noise, based

                                  ///< on which a DC removal filter is used

    int lsps;                     ///< number of LSPs per frame [10 or 16]

    int lsp_q_mode;               ///< defines quantizer defaults [0, 1]

    int lsp_def_mode;             ///< defines different sets of LSP defaults

                                  ///< [0, 1]

    int frame_lsp_bitsize;        ///< size (in bits) of LSPs, when encoded

                                  ///< per-frame (independent coding)

    int sframe_lsp_bitsize;       ///< size (in bits) of LSPs, when encoded

                                  ///< per superframe (residual coding)

    int min_pitch_val;            ///< base value for pitch parsing code

    int max_pitch_val;            ///< max value + 1 for pitch parsing

    int pitch_nbits;              ///< number of bits used to specify the

                                  ///< pitch value in the frame header

    int block_pitch_nbits;        ///< number of bits used to specify the

                                  ///< first block's pitch value

    int block_pitch_range;        ///< range of the block pitch

    int block_delta_pitch_nbits;  ///< number of bits used to specify the

                                  ///< delta pitch between this and the last

                                  ///< block's pitch value, used in all but

                                  ///< first block

    int block_delta_pitch_hrange; ///< 1/2 range of the delta (full range is

                                  ///< from -this to +this-1)

    uint16_t block_conv_table[4]; ///< boundaries for block pitch unit/scale

                                  ///< conversion

    /**

     * @}

     * @defgroup struct_packet Packet values

     * Packet values, specified in the packet header or related to a packet.

     * A packet is considered to be a single unit of data provided to this

     * decoder by the demuxer.

     * @{

     */

    int spillover_nbits;          ///< number of bits of the previous packet's

                                  ///< last superframe preceeding this

                                  ///< packet's first full superframe (useful

                                  ///< for re-synchronization also)

    int has_residual_lsps;        ///< if set, superframes contain one set of

                                  ///< LSPs that cover all frames, encoded as

                                  ///< independent and residual LSPs; if not

                                  ///< set, each frame contains its own, fully

                                  ///< independent, LSPs

    int skip_bits_next;           ///< number of bits to skip at the next call

                                  ///< to #wmavoice_decode_packet() (since

                                  ///< they're part of the previous superframe)

    uint8_t sframe_cache[SFRAME_CACHE_MAXSIZE + FF_INPUT_BUFFER_PADDING_SIZE];

                                  ///< cache for superframe data split over

                                  ///< multiple packets

    int sframe_cache_size;        ///< set to >0 if we have data from an

                                  ///< (incomplete) superframe from a previous

                                  ///< packet that spilled over in the current

                                  ///< packet; specifies the amount of bits in

                                  ///< #sframe_cache

    PutBitContext pb;             ///< bitstream writer for #sframe_cache

    /**

     * @}

     * @defgroup struct_frame Frame and superframe values

     * Superframe and frame data - these can change from frame to frame,

     * although some of them do in that case serve as a cache / history for

     * the next frame or superframe.

     * @{

     */

    double prev_lsps[MAX_LSPS];   ///< LSPs of the last frame of the previous

                                  ///< superframe

    int last_pitch_val;           ///< pitch value of the previous frame

    int last_acb_type;            ///< frame type [0-2] of the previous frame

    int pitch_diff_sh16;          ///< ((cur_pitch_val - #last_pitch_val)

                                  ///< << 16) / #MAX_FRAMESIZE

    float silence_gain;           ///< set for use in blocks if #ACB_TYPE_NONE

    int aw_idx_is_ext;            ///< whether the AW index was encoded in

                                  ///< 8 bits (instead of 6)

    int aw_pulse_range;           ///< the range over which #aw_pulse_set1()

                                  ///< can apply the pulse, relative to the

                                  ///< value in aw_first_pulse_off. The exact

                                  ///< position of the first AW-pulse is within

                                  ///< [pulse_off, pulse_off + this], and

                                  ///< depends on bitstream values; [16 or 24]

    int aw_n_pulses[2];           ///< number of AW-pulses in each block; note

                                  ///< that this number can be negative (in

                                  ///< which case it basically means "zero")

    int aw_first_pulse_off[2];    ///< index of first sample to which to

                                  ///< apply AW-pulses, or -0xff if unset

    int aw_next_pulse_off_cache;  ///< the position (relative to start of the

                                  ///< second block) at which pulses should

                                  ///< start to be positioned, serves as a

                                  ///< cache for pitch-adaptive window pulses

                                  ///< between blocks

    int frame_cntr;               ///< current frame index [0 - 0xFFFE]; is

                                  ///< only used for comfort noise in #pRNG()

    float gain_pred_err[6];       ///< cache for gain prediction

    float excitation_history[MAX_SIGNAL_HISTORY];

                                  ///< cache of the signal of previous

                                  ///< superframes, used as a history for

                                  ///< signal generation

    float synth_history[MAX_LSPS]; ///< see #excitation_history

    /**

     * @}

     * @defgroup post_filter Postfilter values

     * Variables used for postfilter implementation, mostly history for

     * smoothing and so on, and context variables for FFT/iFFT.

     * @{

     */

    RDFTContext rdft, irdft;      ///< contexts for FFT-calculation in the

                                  ///< postfilter (for denoise filter)

    DCTContext dct, dst;          ///< contexts for phase shift (in Hilbert

                                  ///< transform, part of postfilter)

    float sin[511], cos[511];     ///< 8-bit cosine/sine windows over [-pi,pi]

                                  ///< range

    float postfilter_agc;         ///< gain control memory, used in

                                  ///< #adaptive_gain_control()

    float dcf_mem[2];             ///< DC filter history

    float zero_exc_pf[MAX_SIGNAL_HISTORY + MAX_SFRAMESIZE];

                                  ///< zero filter output (i.e. excitation)

                                  ///< by postfilter

    float denoise_filter_cache[MAX_FRAMESIZE];

    int   denoise_filter_cache_size; ///< samples in #denoise_filter_cache

    DECLARE_ALIGNED(16, float, tilted_lpcs_pf)[0x80];

                                  ///< aligned buffer for LPC tilting

    DECLARE_ALIGNED(16, float, denoise_coeffs_pf)[0x80];

                                  ///< aligned buffer for denoise coefficients

    DECLARE_ALIGNED(16, float, synth_filter_out_buf)[0x80 + MAX_LSPS_ALIGN16];

                                  ///< aligned buffer for postfilter speech

                                  ///< synthesis

    /**

     * @}

     */

} WMAVoiceContext;

/**

 * Set up the variable bit mode (VBM) tree from container extradata.

 * @param gb bit I/O context.

 *           The bit context (s->gb) should be loaded with byte 23-46 of the

 *           container extradata (i.e. the ones containing the VBM tree).

 * @param vbm_tree pointer to array to which the decoded VBM tree will be

 *                 written.

 * @return 0 on success, <0 on error.

 */

static av_cold int decode_vbmtree(GetBitContext *gb, int8_t vbm_tree[25])

{

    static const uint8_t bits[] = {

         2,  2,  2,  4,  4,  4,

         6,  6,  6,  8,  8,  8,

        10, 10, 10, 12, 12, 12,

        14, 14, 14, 14

    };

    static const uint16_t codes[] = {

          0x0000, 0x0001, 0x0002,        //              00/01/10

          0x000c, 0x000d, 0x000e,        //           11+00/01/10

          0x003c, 0x003d, 0x003e,        //         1111+00/01/10

          0x00fc, 0x00fd, 0x00fe,        //       111111+00/01/10

          0x03fc, 0x03fd, 0x03fe,        //     11111111+00/01/10

          0x0ffc, 0x0ffd, 0x0ffe,        //   1111111111+00/01/10

          0x3ffc, 0x3ffd, 0x3ffe, 0x3fff // 111111111111+xx

    };

    int cntr[8], n, res;

    memset(vbm_tree, 0xff, sizeof(vbm_tree));

    memset(cntr,     0,    sizeof(cntr));

    for (n = 0; n < 17; n++) {

        res = get_bits(gb, 3);

        if (cntr[res] > 3) // should be >= 3 + (res == 7))

            return -1;

        vbm_tree[res * 3 + cntr[res]++] = n;

    }

    INIT_VLC_STATIC(&frame_type_vlc, VLC_NBITS, sizeof(bits),

                    bits, 1, 1, codes, 2, 2, 132);

    return 0;

}

/**

 * Set up decoder with parameters from demuxer (extradata etc.).

 */

static av_cold int wmavoice_decode_init(AVCodecContext *ctx)

{

    int n, flags, pitch_range, lsp16_flag;

    WMAVoiceContext *s = ctx->priv_data;

    /**

     * Extradata layout:

     * - byte  0-18: WMAPro-in-WMAVoice extradata (see wmaprodec.c),

     * - byte 19-22: flags field (annoyingly in LE; see below for known

     *               values),

     * - byte 23-46: variable bitmode tree (really just 17 * 3 bits,

     *               rest is 0).

     */

    if (ctx->extradata_size != 46) {

        av_log(ctx, AV_LOG_ERROR,

               "Invalid extradata size %d (should be 46)\n",

               ctx->extradata_size);

        return -1;

    }

    flags                = AV_RL32(ctx->extradata + 18);

    s->spillover_bitsize = 3 + av_ceil_log2(ctx->block_align);

    s->do_apf            =    flags & 0x1;

    if (s->do_apf) {

        ff_rdft_init(&s->rdft,  7, DFT_R2C);

        ff_rdft_init(&s->irdft, 7, IDFT_C2R);

        ff_dct_init(&s->dct,  6, DCT_I);

        ff_dct_init(&s->dst,  6, DST_I);

        ff_sine_window_init(s->cos, 256);

        memcpy(&s->sin[255], s->cos, 256 * sizeof(s->cos[0]));

        for (n = 0; n < 255; n++) {

            s->sin[n]       = -s->sin[510 - n];

            s->cos[510 - n] =  s->cos[n];

        }

    }

    s->denoise_strength  =   (flags >> 2) & 0xF;

    if (s->denoise_strength >= 12) {

        av_log(ctx, AV_LOG_ERROR,

               "Invalid denoise filter strength %d (max=11)\n",

               s->denoise_strength);

        return -1;

    }

    s->denoise_tilt_corr = !!(flags & 0x40);

    s->dc_level          =   (flags >> 7) & 0xF;

    s->lsp_q_mode        = !!(flags & 0x2000);

    s->lsp_def_mode      = !!(flags & 0x4000);

    lsp16_flag           =    flags & 0x1000;

    if (lsp16_flag) {

        s->lsps               = 16;

        s->frame_lsp_bitsize  = 34;

        s->sframe_lsp_bitsize = 60;

    } else {

        s->lsps               = 10;

        s->frame_lsp_bitsize  = 24;

        s->sframe_lsp_bitsize = 48;

    }

    for (n = 0; n < s->lsps; n++)

        s->prev_lsps[n] = M_PI * (n + 1.0) / (s->lsps + 1.0);

    init_get_bits(&s->gb, ctx->extradata + 22, (ctx->extradata_size - 22) << 3);

    if (decode_vbmtree(&s->gb, s->vbm_tree) < 0) {

        av_log(ctx, AV_LOG_ERROR, "Invalid VBM tree; broken extradata?\n");

        return -1;

    }

    s->min_pitch_val    = ((ctx->sample_rate << 8)      /  400 + 50) >> 8;

    s->max_pitch_val    = ((ctx->sample_rate << 8) * 37 / 2000 + 50) >> 8;

    pitch_range         = s->max_pitch_val - s->min_pitch_val;

    s->pitch_nbits      = av_ceil_log2(pitch_range);

    s->last_pitch_val   = 40;

    s->last_acb_type    = ACB_TYPE_NONE;

    s->history_nsamples = s->max_pitch_val + 8;

    if (s->min_pitch_val < 1 || s->history_nsamples > MAX_SIGNAL_HISTORY) {

        int min_sr = ((((1 << 8) - 50) * 400) + 0xFF) >> 8,

            max_sr = ((((MAX_SIGNAL_HISTORY - 8) << 8) + 205) * 2000 / 37) >> 8;

        av_log(ctx, AV_LOG_ERROR,

               "Unsupported samplerate %d (min=%d, max=%d)\n",

               ctx->sample_rate, min_sr, max_sr); // 322-22097 Hz

        return -1;

    }

    s->block_conv_table[0]      = s->min_pitch_val;

    s->block_conv_table[1]      = (pitch_range * 25) >> 6;

    s->block_conv_table[2]      = (pitch_range * 44) >> 6;

    s->block_conv_table[3]      = s->max_pitch_val - 1;

    s->block_delta_pitch_hrange = (pitch_range >> 3) & ~0xF;

    s->block_delta_pitch_nbits  = 1 + av_ceil_log2(s->block_delta_pitch_hrange);

    s->block_pitch_range        = s->block_conv_table[2] +

                                  s->block_conv_table[3] + 1 +

                                  2 * (s->block_conv_table[1] - 2 * s->min_pitch_val);

    s->block_pitch_nbits        = av_ceil_log2(s->block_pitch_range);

    ctx->sample_fmt             = AV_SAMPLE_FMT_FLT;

    return 0;

}

/**

 * @defgroup postfilter Postfilter functions

 * Postfilter functions (gain control, wiener denoise filter, DC filter,

 * kalman smoothening, plus surrounding code to wrap it)

 * @{

 */

/**

 * Adaptive gain control (as used in postfilter).

 *

 * Identical to #ff_adaptive_gain_control() in acelp_vectors.c, except

 * that the energy here is calculated using sum(abs(...)), whereas the

 * other codecs (e.g. AMR-NB, SIPRO) use sqrt(dotproduct(...)).

 *

 * @param out output buffer for filtered samples

 * @param in input buffer containing the samples as they are after the

 *           postfilter steps so far

 * @param speech_synth input buffer containing speech synth before postfilter

 * @param size input buffer size

 * @param alpha exponential filter factor

 * @param gain_mem pointer to filter memory (single float)

 */

static void adaptive_gain_control(float *out, const float *in,

                                  const float *speech_synth,

                                  int size, float alpha, float *gain_mem)

{

    int i;

    float speech_energy = 0.0, postfilter_energy = 0.0, gain_scale_factor;

    float mem = *gain_mem;

    for (i = 0; i < size; i++) {

        speech_energy     += fabsf(speech_synth[i]);

        postfilter_energy += fabsf(in[i]);

    }

    gain_scale_factor = (1.0 - alpha) * speech_energy / postfilter_energy;

    for (i = 0; i < size; i++) {

        mem = alpha * mem + gain_scale_factor;

        out[i] = in[i] * mem;

    }

    *gain_mem = mem;

}

/**

 * Kalman smoothing function.

 *

 * This function looks back pitch +/- 3 samples back into history to find

 * the best fitting curve (that one giving the optimal gain of the two

 * signals, i.e. the highest dot product between the two), and then

 * uses that signal history to smoothen the output of the speech synthesis

 * filter.

 *

 * @param s WMA Voice decoding context

 * @param pitch pitch of the speech signal

 * @param in input speech signal

 * @param out output pointer for smoothened signal

 * @param size input/output buffer size

 *

 * @returns -1 if no smoothening took place, e.g. because no optimal

 *          fit could be found, or 0 on success.

 */

static int kalman_smoothen(WMAVoiceContext *s, int pitch,

                           const float *in, float *out, int size)

{

    int n;

    float optimal_gain = 0, dot;

    const float *ptr = &in[-FFMAX(s->min_pitch_val, pitch - 3)],

                *end = &in[-FFMIN(s->max_pitch_val, pitch + 3)],

                *best_hist_ptr;

    /* find best fitting point in history */

    do {

        dot = ff_dot_productf(in, ptr, size);

        if (dot > optimal_gain) {

            optimal_gain  = dot;

            best_hist_ptr = ptr;

        }

    } while (--ptr >= end);

    if (optimal_gain <= 0)

        return -1;

    dot = ff_dot_productf(best_hist_ptr, best_hist_ptr, size);

    if (dot <= 0) // would be 1.0

        return -1;

    if (optimal_gain <= dot) {

        dot = dot / (dot + 0.6 * optimal_gain); // 0.625-1.000

    } else

        dot = 0.625;

    /* actual smoothing */

    for (n = 0; n < size; n++)

        out[n] = best_hist_ptr[n] + dot * (in[n] - best_hist_ptr[n]);

    return 0;

}

/**

 * Get the tilt factor of a formant filter from its transfer function

 * @see #tilt_factor() in amrnbdec.c, which does essentially the same,

 *      but somehow (??) it does a speech synthesis filter in the

 *      middle, which is missing here

 *

 * @param lpcs LPC coefficients

 * @param n_lpcs Size of LPC buffer

 * @returns the tilt factor

 */

static float tilt_factor(const float *lpcs, int n_lpcs)

{

    float rh0, rh1;

    rh0 = 1.0     + ff_dot_productf(lpcs,  lpcs,    n_lpcs);

    rh1 = lpcs[0] + ff_dot_productf(lpcs, &lpcs[1], n_lpcs - 1);

    return rh1 / rh0;

}

/**

 * Derive denoise filter coefficients (in real domain) from the LPCs.

 */

static void calc_input_response(WMAVoiceContext *s, float *lpcs,

                                int fcb_type, float *coeffs, int remainder)

{

    float last_coeff, min = 15.0, max = -15.0;

    float irange, angle_mul, gain_mul, range, sq;

    int n, idx;

    /* Create frequency power spectrum of speech input (i.e. RDFT of LPCs) */

    ff_rdft_calc(&s->rdft, lpcs);

#define log_range(var, assign) do { \

        float tmp = log10f(assign);  var = tmp; \

        max       = FFMAX(max, tmp); min = FFMIN(min, tmp); \

    } while (0)

    log_range(last_coeff,  lpcs[1]         * lpcs[1]);

    for (n = 1; n < 64; n++)

        log_range(lpcs[n], lpcs[n * 2]     * lpcs[n * 2] +

                           lpcs[n * 2 + 1] * lpcs[n * 2 + 1]);

    log_range(lpcs[0],     lpcs[0]         * lpcs[0]);

#undef log_range

    range    = max - min;

    lpcs[64] = last_coeff;

    /* Now, use this spectrum to pick out these frequencies with higher

     * (relative) power/energy (which we then take to be "not noise"),

     * and set up a table (still in lpc[]) of (relative) gains per frequency.

     * These frequencies will be maintained, while others ("noise") will be

     * decreased in the filter output. */

    irange    = 64.0 / range; // so irange*(max-value) is in the range [0, 63]

    gain_mul  = range * (fcb_type == FCB_TYPE_HARDCODED ? (5.0 / 13.0) :

                                                          (5.0 / 14.7));

    angle_mul = gain_mul * (8.0 * M_LN10 / M_PI);

    for (n = 0; n <= 64; n++) {

        float pwr;

        idx = FFMAX(0, lrint((max - lpcs[n]) * irange) - 1);

        pwr = wmavoice_denoise_power_table[s->denoise_strength][idx];

        lpcs[n] = angle_mul * pwr;

        /* 70.57 =~ 1/log10(1.0331663) */

        idx = (pwr * gain_mul - 0.0295) * 70.570526123;

        if (idx > 127) { // fallback if index falls outside table range

            coeffs[n] = wmavoice_energy_table[127] *

                        powf(1.0331663, idx - 127);

        } else

            coeffs[n] = wmavoice_energy_table[FFMAX(0, idx)];

    }

    /* calculate the Hilbert transform of the gains, which we do (since this

     * is a sinus input) by doing a phase shift (in theory, H(sin())=cos()).

     * Hilbert_Transform(RDFT(x)) = Laplace_Transform(x), which calculates the

     * "moment" of the LPCs in this filter. */

    ff_dct_calc(&s->dct, lpcs);

    ff_dct_calc(&s->dst, lpcs);

    /* Split out the coefficient indexes into phase/magnitude pairs */

    idx = 255 + av_clip(lpcs[64],               -255, 255);

    coeffs[0]  = coeffs[0]  * s->cos[idx];

    idx = 255 + av_clip(lpcs[64] - 2 * lpcs[63], -255, 255);

    last_coeff = coeffs[64] * s->cos[idx];

    for (n = 63;; n--) {

        idx = 255 + av_clip(-lpcs[64] - 2 * lpcs[n - 1], -255, 255);

        coeffs[n * 2 + 1] = coeffs[n] * s->sin[idx];

        coeffs[n * 2]     = coeffs[n] * s->cos[idx];

        if (!--n) break;

        idx = 255 + av_clip( lpcs[64] - 2 * lpcs[n - 1], -255, 255);

        coeffs[n * 2 + 1] = coeffs[n] * s->sin[idx];

        coeffs[n * 2]     = coeffs[n] * s->cos[idx];

    }

    coeffs[1] = last_coeff;

    /* move into real domain */

    ff_rdft_calc(&s->irdft, coeffs);

    /* tilt correction and normalize scale */

    memset(&coeffs[remainder], 0, sizeof(coeffs[0]) * (128 - remainder));

    if (s->denoise_tilt_corr) {

        float tilt_mem = 0;

        coeffs[remainder - 1] = 0;

        ff_tilt_compensation(&tilt_mem,

                             -1.8 * tilt_factor(coeffs, remainder - 1),

                             coeffs, remainder);

    }

    sq = (1.0 / 64.0) * sqrtf(1 / ff_dot_productf(coeffs, coeffs, remainder));

    for (n = 0; n < remainder; n++)

        coeffs[n] *= sq;

}

/**

 * This function applies a Wiener filter on the (noisy) speech signal as

 * a means to denoise it.

 *

 * - take RDFT of LPCs to get the power spectrum of the noise + speech;

 * - using this power spectrum, calculate (for each frequency) the Wiener

 *    filter gain, which depends on the frequency power and desired level

 *    of noise subtraction (when set too high, this leads to artifacts)

 *    We can do this symmetrically over the X-axis (so 0-4kHz is the inverse

 *    of 4-8kHz);

 * - by doing a phase shift, calculate the Hilbert transform of this array

 *    of per-frequency filter-gains to get the filtering coefficients;

 * - smoothen/normalize/de-tilt these filter coefficients as desired;

 * - take RDFT of noisy sound, apply the coefficients and take its IRDFT

 *    to get the denoised speech signal;

 * - the leftover (i.e. output of the IRDFT on denoised speech data beyond

 *    the frame boundary) are saved and applied to subsequent frames by an

 *    overlap-add method (otherwise you get clicking-artifacts).

 *

 * @param s WMA Voice decoding context

 * @param fcb_type Frame (codebook) type

 * @param synth_pf input: the noisy speech signal, output: denoised speech

 *                 data; should be 16-byte aligned (for ASM purposes)

 * @param size size of the speech data

 * @param lpcs LPCs used to synthesize this frame's speech data

 */

static void wiener_denoise(WMAVoiceContext *s, int fcb_type,

                           float *synth_pf, int size,

                           const float *lpcs)

{

    int remainder, lim, n;

    if (fcb_type != FCB_TYPE_SILENCE) {

        float *tilted_lpcs = s->tilted_lpcs_pf,

              *coeffs = s->denoise_coeffs_pf, tilt_mem = 0;

        tilted_lpcs[0]           = 1.0;

        memcpy(&tilted_lpcs[1], lpcs, sizeof(lpcs[0]) * s->lsps);

        memset(&tilted_lpcs[s->lsps + 1], 0,

               sizeof(tilted_lpcs[0]) * (128 - s->lsps - 1));

        ff_tilt_compensation(&tilt_mem, 0.7 * tilt_factor(lpcs, s->lsps),

                             tilted_lpcs, s->lsps + 2);

        /* The IRDFT output (127 samples for 7-bit filter) beyond the frame

         * size is applied to the next frame. All input beyond this is zero,

         * and thus all output beyond this will go towards zero, hence we can

         * limit to min(size-1, 127-size) as a performance consideration. */

        remainder = FFMIN(127 - size, size - 1);

        calc_input_response(s, tilted_lpcs, fcb_type, coeffs, remainder);

        /* apply coefficients (in frequency spectrum domain), i.e. complex

         * number multiplication */

        memset(&synth_pf[size], 0, sizeof(synth_pf[0]) * (128 - size));

        ff_rdft_calc(&s->rdft, synth_pf);

        ff_rdft_calc(&s->rdft, coeffs);

        synth_pf[0] *= coeffs[0];

        synth_pf[1] *= coeffs[1];

        for (n = 1; n < 64; n++) {

            float v1 = synth_pf[n * 2], v2 = synth_pf[n * 2 + 1];

            synth_pf[n * 2]     = v1 * coeffs[n * 2] - v2 * coeffs[n * 2 + 1];

            synth_pf[n * 2 + 1] = v2 * coeffs[n * 2] + v1 * coeffs[n * 2 + 1];

        }

        ff_rdft_calc(&s->irdft, synth_pf);

    }

    /* merge filter output with the history of previous runs */

    if (s->denoise_filter_cache_size) {

        lim = FFMIN(s->denoise_filter_cache_size, size);

        for (n = 0; n < lim; n++)

            synth_pf[n] += s->denoise_filter_cache[n];

        s->denoise_filter_cache_size -= lim;

        memmove(s->denoise_filter_cache, &s->denoise_filter_cache[size],

                sizeof(s->denoise_filter_cache[0]) * s->denoise_filter_cache_size);

    }

    /* move remainder of filter output into a cache for future runs */

    if (fcb_type != FCB_TYPE_SILENCE) {

        lim = FFMIN(remainder, s->denoise_filter_cache_size);

        for (n = 0; n < lim; n++)

            s->denoise_filter_cache[n] += synth_pf[size + n];

        if (lim < remainder) {

            memcpy(&s->denoise_filter_cache[lim], &synth_pf[size + lim],

                   sizeof(s->denoise_filter_cache[0]) * (remainder - lim));

            s->denoise_filter_cache_size = remainder;

        }

    }

}

/**

 * Averaging projection filter, the postfilter used in WMAVoice.

 *

 * This uses the following steps:

 * - A zero-synthesis filter (generate excitation from synth signal)

 * - Kalman smoothing on excitation, based on pitch

 * - Re-synthesized smoothened output

 * - Iterative Wiener denoise filter

 * - Adaptive gain filter

 * - DC filter

 *

 * @param s WMAVoice decoding context

 * @param synth Speech synthesis output (before postfilter)

 * @param samples Output buffer for filtered samples

 * @param size Buffer size of synth & samples

 * @param lpcs Generated LPCs used for speech synthesis

 * @param zero_exc_pf destination for zero synthesis filter (16-byte aligned)

 * @param fcb_type Frame type (silence, hardcoded, AW-pulses or FCB-pulses)

 * @param pitch Pitch of the input signal

 */

static void postfilter(WMAVoiceContext *s, const float *synth,

                       float *samples,    int size,

                       const float *lpcs, float *zero_exc_pf,

                       int fcb_type,      int pitch)

{

    float synth_filter_in_buf[MAX_FRAMESIZE / 2],

          *synth_pf = &s->synth_filter_out_buf[MAX_LSPS_ALIGN16],

          *synth_filter_in = zero_exc_pf;

    assert(size <= MAX_FRAMESIZE / 2);

    /* generate excitation from input signal */

    ff_celp_lp_zero_synthesis_filterf(zero_exc_pf, lpcs, synth, size, s->lsps);

    if (fcb_type >= FCB_TYPE_AW_PULSES &&

        !kalman_smoothen(s, pitch, zero_exc_pf, synth_filter_in_buf, size))

        synth_filter_in = synth_filter_in_buf;

    /* re-synthesize speech after smoothening, and keep history */

    ff_celp_lp_synthesis_filterf(synth_pf, lpcs,

                                 synth_filter_in, size, s->lsps);

    memcpy(&synth_pf[-s->lsps], &synth_pf[size - s->lsps],

           sizeof(synth_pf[0]) * s->lsps);

    wiener_denoise(s, fcb_type, synth_pf, size, lpcs);

    adaptive_gain_control(samples, synth_pf, synth, size, 0.99,

                          &s->postfilter_agc);

    if (s->dc_level > 8) {

        /* remove ultra-low frequency DC noise / highpass filter;

         * coefficients are identical to those used in SIPR decoding,

         * and very closely resemble those used in AMR-NB decoding. */

        ff_acelp_apply_order_2_transfer_function(samples, samples,

            (const float[2]) { -1.99997,      1.0 },

            (const float[2]) { -1.9330735188, 0.93589198496 },

            0.93980580475, s->dcf_mem, size);

    }

}

/**

 * @}

 */

/**

 * Dequantize LSPs

 * @param lsps output pointer to the array that will hold the LSPs

 * @param num number of LSPs to be dequantized

 * @param values quantized values, contains n_stages values

 * @param sizes range (i.e. max value) of each quantized value

 * @param n_stages number of dequantization runs

 * @param table dequantization table to be used

 * @param mul_q LSF multiplier

 * @param base_q base (lowest) LSF values

 */

static void dequant_lsps(double *lsps, int num,

                         const uint16_t *values,

                         const uint16_t *sizes,

                         int n_stages, const uint8_t *table,

                         const double *mul_q,

                         const double *base_q)

{

    int n, m;

    memset(lsps, 0, num * sizeof(*lsps));

    for (n = 0; n < n_stages; n++) {

        const uint8_t *t_off = &table[values[n] * num];

        double base = base_q[n], mul = mul_q[n];

        for (m = 0; m < num; m++)

            lsps[m] += base + mul * t_off[m];

        table += sizes[n] * num;

    }

}

/**

 * @defgroup lsp_dequant LSP dequantization routines

 * LSP dequantization routines, for 10/16LSPs and independent/residual coding.

 * @note we assume enough bits are available, caller should check.

 * lsp10i() consumes 24 bits; lsp10r() consumes an additional 24 bits;

 * lsp16i() consumes 34 bits; lsp16r() consumes an additional 26 bits.

 * @{

 */

/**

 * Parse 10 independently-coded LSPs.

 */

static void dequant_lsp10i(GetBitContext *gb, double *lsps)

{

    static const uint16_t vec_sizes[4] = { 256, 64, 32, 32 };

    static const double mul_lsf[4] = {

        5.2187144800e-3,    1.4626986422e-3,

        9.6179549166e-4,    1.1325736225e-3

    };

    static const double base_lsf[4] = {

        M_PI * -2.15522e-1, M_PI * -6.1646e-2,

        M_PI * -3.3486e-2,  M_PI * -5.7408e-2

    };

    uint16_t v[4];

    v[0] = get_bits(gb, 8);

    v[1] = get_bits(gb, 6);

    v[2] = get_bits(gb, 5);

    v[3] = get_bits(gb, 5);

    dequant_lsps(lsps, 10, v, vec_sizes, 4, wmavoice_dq_lsp10i,

                 mul_lsf, base_lsf);

}

/**

 * Parse 10 independently-coded LSPs, and then derive the tables to

 * generate LSPs for the other frames from them (residual coding).

 */

static void dequant_lsp10r(GetBitContext *gb,

                           double *i_lsps, const double *old,

                           double *a1, double *a2, int q_mode)

{

    static const uint16_t vec_sizes[3] = { 128, 64, 64 };

    static const double mul_lsf[3] = {

        2.5807601174e-3,    1.2354460219e-3,   1.1763821673e-3

    };

    static const double base_lsf[3] = {

        M_PI * -1.07448e-1, M_PI * -5.2706e-2, M_PI * -5.1634e-2

    };

    const float (*ipol_tab)[2][10] = q_mode ?

        wmavoice_lsp10_intercoeff_b : wmavoice_lsp10_intercoeff_a;

    uint16_t interpol, v[3];

    int n;

    dequant_lsp10i(gb, i_lsps);

    interpol = get_bits(gb, 5);

    v[0]     = get_bits(gb, 7);

    v[1]     = get_bits(gb, 6);

    v[2]     = get_bits(gb, 6);

    for (n = 0; n < 10; n++) {

        double delta = old[n] - i_lsps[n];

        a1[n]        = ipol_tab[interpol][0][n] * delta + i_lsps[n];

        a1[10 + n]   = ipol_tab[interpol][1][n] * delta + i_lsps[n];

    }

    dequant_lsps(a2, 20, v, vec_sizes, 3, wmavoice_dq_lsp10r,

                 mul_lsf, base_lsf);

}

/**

 * Parse 16 independently-coded LSPs.

 */

static void dequant_lsp16i(GetBitContext *gb, double *lsps)

{

    static const uint16_t vec_sizes[5] = { 256, 64, 128, 64, 128 };

    static const double mul_lsf[5] = {

        3.3439586280e-3,    6.9908173703e-4,

        3.3216608306e-3,    1.0334960326e-3,

        3.1899104283e-3

    };

    static const double base_lsf[5] = {

        M_PI * -1.27576e-1, M_PI * -2.4292e-2,

        M_PI * -1.28094e-1, M_PI * -3.2128e-2,

        M_PI * -1.29816e-1

    };

    uint16_t v[5];

    v[0] = get_bits(gb, 8);

    v[1] = get_bits(gb, 6);

    v[2] = get_bits(gb, 7);

    v[3] = get_bits(gb, 6);

    v[4] = get_bits(gb, 7);

    dequant_lsps( lsps,     5,  v,     vec_sizes,    2,

                 wmavoice_dq_lsp16i1,  mul_lsf,     base_lsf);

    dequant_lsps(&lsps[5],  5, &v[2], &vec_sizes[2], 2,

                 wmavoice_dq_lsp16i2, &mul_lsf[2], &base_lsf[2]);

    dequant_lsps(&lsps[10], 6, &v[4], &vec_sizes[4], 1,

                 wmavoice_dq_lsp16i3, &mul_lsf[4], &base_lsf[4]);

}

/**

 * Parse 16 independently-coded LSPs, and then derive the tables to

 * generate LSPs for the other frames from them (residual coding).

 */

static void dequant_lsp16r(GetBitContext *gb,

                           double *i_lsps, const double *old,

                           double *a1, double *a2, int q_mode)

{

    static const uint16_t vec_sizes[3] = { 128, 128, 128 };

    static const double mul_lsf[3] = {

        1.2232979501e-3,   1.4062241527e-3,   1.6114744851e-3

    };

    static const double base_lsf[3] = {

        M_PI * -5.5830e-2, M_PI * -5.2908e-2, M_PI * -5.4776e-2

    };

    const float (*ipol_tab)[2][16] = q_mode ?

        wmavoice_lsp16_intercoeff_b : wmavoice_lsp16_intercoeff_a;

    uint16_t interpol, v[3];

    int n;

    dequant_lsp16i(gb, i_lsps);

    interpol = get_bits(gb, 5);

    v[0]     = get_bits(gb, 7);

    v[1]     = get_bits(gb, 7);

    v[2]     = get_bits(gb, 7);

    for (n = 0; n < 16; n++) {

        double delta = old[n] - i_lsps[n];

        a1[n]        = ipol_tab[interpol][0][n] * delta + i_lsps[n];

        a1[16 + n]   = ipol_tab[interpol][1][n] * delta + i_lsps[n];

    }

    dequant_lsps( a2,     10,  v,     vec_sizes,    1,

                 wmavoice_dq_lsp16r1,  mul_lsf,     base_lsf);

    dequant_lsps(&a2[10], 10, &v[1], &vec_sizes[1], 1,

                 wmavoice_dq_lsp16r2, &mul_lsf[1], &base_lsf[1]);

    dequant_lsps(&a2[20], 12, &v[2], &vec_sizes[2], 1,

                 wmavoice_dq_lsp16r3, &mul_lsf[2], &base_lsf[2]);

}

/**

 * @}

 * @defgroup aw Pitch-adaptive window coding functions

 * The next few functions are for pitch-adaptive window coding.

 * @{

 */

/**

 * Parse the offset of the first pitch-adaptive window pulses, and

 * the distribution of pulses between the two blocks in this frame.

 * @param s WMA Voice decoding context private data

 * @param gb bit I/O context

 * @param pitch pitch for each block in this frame

 */

static void aw_parse_coords(WMAVoiceContext *s, GetBitContext *gb,

                            const int *pitch)

{

    static const int16_t start_offset[94] = {

        -11,  -9,  -7,  -5,  -3,  -1,   1,   3,   5,   7,   9,  11,

         13,  15,  18,  17,  19,  20,  21,  22,  23,  24,  25,  26,

         27,  28,  29,  30,  31,  32,  33,  35,  37,  39,  41,  43,

         45,  47,  49,  51,  53,  55,  57,  59,  61,  63,  65,  67,

         69,  71,  73,  75,  77,  79,  81,  83,  85,  87,  89,  91,

         93,  95,  97,  99, 101, 103, 105, 107, 109, 111, 113, 115,

        117, 119, 121, 123, 125, 127, 129, 131, 133, 135, 137, 139,

        141, 143, 145, 147, 149, 151, 153, 155, 157, 159

    };

    int bits, offset;

    /* position of pulse */

    s->aw_idx_is_ext = 0;

    if ((bits = get_bits(gb, 6)) >= 54) {

        s->aw_idx_is_ext = 1;

        bits += (bits - 54) * 3 + get_bits(gb, 2);

    }

    /* for a repeated pulse at pulse_off with a pitch_lag of pitch[], count

     * the distribution of the pulses in each block contained in this frame. */

    s->aw_pulse_range        = FFMIN(pitch[0], pitch[1]) > 32 ? 24 : 16;

    for (offset = start_offset[bits]; offset < 0; offset += pitch[0]) ;

    s->aw_n_pulses[0]        = (pitch[0] - 1 + MAX_FRAMESIZE / 2 - offset) / pitch[0];

    s->aw_first_pulse_off[0] = offset - s->aw_pulse_range / 2;

    offset                  += s->aw_n_pulses[0] * pitch[0];

    s->aw_n_pulses[1]        = (pitch[1] - 1 + MAX_FRAMESIZE - offset) / pitch[1];

    s->aw_first_pulse_off[1] = offset - (MAX_FRAMESIZE + s->aw_pulse_range) / 2;

    /* if continuing from a position before the block, reset position to

     * start of block (when corrected for the range over which it can be

     * spread in aw_pulse_set1()). */

    if (start_offset[bits] < MAX_FRAMESIZE / 2) {

        while (s->aw_first_pulse_off[1] - pitch[1] + s->aw_pulse_range > 0)

            s->aw_first_pulse_off[1] -= pitch[1];

        if (start_offset[bits] < 0)

            while (s->aw_first_pulse_off[0] - pitch[0] + s->aw_pulse_range > 0)

                s->aw_first_pulse_off[0] -= pitch[0];

    }

}

/**

 * Apply second set of pitch-adaptive window pulses.

 * @param s WMA Voice decoding context private data

 * @param gb bit I/O context

 * @param block_idx block index in frame [0, 1]

 * @param fcb structure containing fixed codebook vector info

 */

static void aw_pulse_set2(WMAVoiceContext *s, GetBitContext *gb,

                          int block_idx, AMRFixed *fcb)

{

    uint16_t use_mask_mem[9]; // only 5 are used, rest is padding

    uint16_t *use_mask = use_mask_mem + 2;

    /* in this function, idx is the index in the 80-bit (+ padding) use_mask

     * bit-array. Since use_mask consists of 16-bit values, the lower 4 bits

     * of idx are the position of the bit within a particular item in the

     * array (0 being the most significant bit, and 15 being the least

     * significant bit), and the remainder (>> 4) is the index in the

     * use_mask[]-array. This is faster and uses less memory than using a

     * 80-byte/80-int array. */

    int pulse_off = s->aw_first_pulse_off[block_idx],

        pulse_start, n, idx, range, aidx, start_off = 0;

    /* set offset of first pulse to within this block */

    if (s->aw_n_pulses[block_idx] > 0)

        while (pulse_off + s->aw_pulse_range < 1)

            pulse_off += fcb->pitch_lag;

    /* find range per pulse */

    if (s->aw_n_pulses[0] > 0) {

        if (block_idx == 0) {

            range = 32;

        } else /* block_idx = 1 */ {

            range = 8;

            if (s->aw_n_pulses[block_idx] > 0)

                pulse_off = s->aw_next_pulse_off_cache;

        }

    } else

        range = 16;

    pulse_start = s->aw_n_pulses[block_idx] > 0 ? pulse_off - range / 2 : 0;

    /* aw_pulse_set1() already applies pulses around pulse_off (to be exactly,

     * in the range of [pulse_off, pulse_off + s->aw_pulse_range], and thus

     * we exclude that range from being pulsed again in this function. */

    memset(&use_mask[-2], 0, 2 * sizeof(use_mask[0]));

    memset( use_mask,   -1, 5 * sizeof(use_mask[0]));

    memset(&use_mask[5], 0, 2 * sizeof(use_mask[0]));

    if (s->aw_n_pulses[block_idx] > 0)

        for (idx = pulse_off; idx < MAX_FRAMESIZE / 2; idx += fcb->pitch_lag) {

            int excl_range         = s->aw_pulse_range; // always 16 or 24

            uint16_t *use_mask_ptr = &use_mask[idx >> 4];

            int first_sh           = 16 - (idx & 15);

            *use_mask_ptr++       &= 0xFFFF << first_sh;

            excl_range            -= first_sh;

            if (excl_range >= 16) {

                *use_mask_ptr++    = 0;

                *use_mask_ptr     &= 0xFFFF >> (excl_range - 16);

            } else

                *use_mask_ptr     &= 0xFFFF >> excl_range;

        }

    /* find the 'aidx'th offset that is not excluded */

    aidx = get_bits(gb, s->aw_n_pulses[0] > 0 ? 5 - 2 * block_idx : 4);

    for (n = 0; n <= aidx; pulse_start++) {

        for (idx = pulse_start; idx < 0; idx += fcb->pitch_lag) ;

        if (idx >= MAX_FRAMESIZE / 2) { // find from zero

            if (use_mask[0])      idx = 0x0F;

            else if (use_mask[1]) idx = 0x1F;

            else if (use_mask[2]) idx = 0x2F;

            else if (use_mask[3]) idx = 0x3F;

            else if (use_mask[4]) idx = 0x4F;

            else                  return;

            idx -= av_log2_16bit(use_mask[idx >> 4]);

        }

        if (use_mask[idx >> 4] & (0x8000 >> (idx & 15))) {

            use_mask[idx >> 4] &= ~(0x8000 >> (idx & 15));

            n++;

            start_off = idx;

        }

    }

    fcb->x[fcb->n] = start_off;

    fcb->y[fcb->n] = get_bits1(gb) ? -1.0 : 1.0;

    fcb->n++;

    /* set offset for next block, relative to start of that block */

    n = (MAX_FRAMESIZE / 2 - start_off) % fcb->pitch_lag;

    s->aw_next_pulse_off_cache = n ? fcb->pitch_lag - n : 0;

}

/**

 * Apply first set of pitch-adaptive window pulses.

 * @param s WMA Voice decoding context private data

 * @param gb bit I/O context

 * @param block_idx block index in frame [0, 1]

 * @param fcb storage location for fixed codebook pulse info

 */

static void aw_pulse_set1(WMAVoiceContext *s, GetBitContext *gb,

                          int block_idx, AMRFixed *fcb)

{

    int val = get_bits(gb, 12 - 2 * (s->aw_idx_is_ext && !block_idx));

    float v;

    if (s->aw_n_pulses[block_idx] > 0) {

        int n, v_mask, i_mask, sh, n_pulses;

        if (s->aw_pulse_range == 24) { // 3 pulses, 1:sign + 3:index each

            n_pulses = 3;

            v_mask   = 8;

            i_mask   = 7;

            sh       = 4;

        } else { // 4 pulses, 1:sign + 2:index each

            n_pulses = 4;

            v_mask   = 4;

            i_mask   = 3;

            sh       = 3;

        }

        for (n = n_pulses - 1; n >= 0; n--, val >>= sh) {

            fcb->y[fcb->n] = (val & v_mask) ? -1.0 : 1.0;

            fcb->x[fcb->n] = (val & i_mask) * n_pulses + n +

                                 s->aw_first_pulse_off[block_idx];

            while (fcb->x[fcb->n] < 0)

                fcb->x[fcb->n] += fcb->pitch_lag;

            if (fcb->x[fcb->n] < MAX_FRAMESIZE / 2)

                fcb->n++;

        }

    } else {

        int num2 = (val & 0x1FF) >> 1, delta, idx;

        if (num2 < 1 * 79)      { delta = 1; idx = num2 + 1; }

        else if (num2 < 2 * 78) { delta = 3; idx = num2 + 1 - 1 * 77; }

        else if (num2 < 3 * 77) { delta = 5; idx = num2 + 1 - 2 * 76; }

        else                    { delta = 7; idx = num2 + 1 - 3 * 75; }

        v = (val & 0x200) ? -1.0 : 1.0;

        fcb->no_repeat_mask |= 3 << fcb->n;

        fcb->x[fcb->n]       = idx - delta;

        fcb->y[fcb->n]       = v;

        fcb->x[fcb->n + 1]   = idx;

        fcb->y[fcb->n + 1]   = (val & 1) ? -v : v;

        fcb->n              += 2;

    }

}

/**

 * @}

 *

 * Generate a random number from frame_cntr and block_idx, which will lief

 * in the range [0, 1000 - block_size] (so it can be used as an index in a

 * table of size 1000 of which you want to read block_size entries).

 *

 * @param frame_cntr current frame number

 * @param block_num current block index

 * @param block_size amount of entries we want to read from a table

 *                   that has 1000 entries

 * @return a (non-)random number in the [0, 1000 - block_size] range.

 */

static int pRNG(int frame_cntr, int block_num, int block_size)

{

    /* array to simplify the calculation of z:

     * y = (x % 9) * 5 + 6;

     * z = (49995 * x) / y;

     * Since y only has 9 values, we can remove the division by using a

     * LUT and using FASTDIV-style divisions. For each of the 9 values

     * of y, we can rewrite z as:

     * z = x * (49995 / y) + x * ((49995 % y) / y)

     * In this table, each col represents one possible value of y, the

     * first number is 49995 / y, and the second is the FASTDIV variant

     * of 49995 % y / y. */

    static const unsigned int div_tbl[9][2] = {

        { 8332,  3 * 715827883U }, // y =  6

        { 4545,  0 * 390451573U }, // y = 11

        { 3124, 11 * 268435456U }, // y = 16

        { 2380, 15 * 204522253U }, // y = 21

        { 1922, 23 * 165191050U }, // y = 26

        { 1612, 23 * 138547333U }, // y = 31

        { 1388, 27 * 119304648U }, // y = 36

        { 1219, 16 * 104755300U }, // y = 41

        { 1086, 39 *  93368855U }  // y = 46

    };

    unsigned int z, y, x = MUL16(block_num, 1877) + frame_cntr;

    if (x >= 0xFFFF) x -= 0xFFFF;   // max value of x is 8*1877+0xFFFE=0x13AA6,

                                    // so this is effectively a modulo (%)

    y = x - 9 * MULH(477218589, x); // x % 9

    z = (uint16_t) (x * div_tbl[y][0] + UMULH(x, div_tbl[y][1]));

                                    // z = x * 49995 / (y * 5 + 6)

    return z % (1000 - block_size);

}

/**

 * Parse hardcoded signal for a single block.

 * @note see #synth_block().

 */

static void synth_block_hardcoded(WMAVoiceContext *s, GetBitContext *gb,

                                 int block_idx, int size,